Machine learning reveals orbital interaction in materials

Tien Lam Pham,Hiori Kino,Kiyoyuki Terakura,Takashi Miyake,Koji Tsuda,Ichigaku Takigawa,Hieu Chi Dam

doi:10.1080/14686996.2017.1378060

Abstract

We propose a novel representation of materials named an ‘orbital-field matrix (OFM)’, which is based on the distribution of valence shell electrons. We demonstrate that this new representation can be highly useful in mining material data. Experimental investigation shows that the formation energies of crystalline materials, atomization energies of molecular materials, and local magnetic moments of the constituent atoms in bimetal alloys of lanthanide metal and transition-metal can be predicted with high accuracy using the OFM. Knowledge regarding the role of the coordination numbers of the transition-metal and lanthanide elements in determining the local magnetic moments of the transition-metal sites can be acquired directly from decision tree regression analyses using the OFM.

Highlights

The increasing volume of available experimental and quantum-computational material data, along with the development of machine learning techniques, has provided a new opportunity to develop methods for accelerating discoveries of new materials and physical and chemical phenomena
We focus on the local magnetic moments of transition-metals in lanthanide metal and transition-metal (LAT) alloys, the dataset of which includes 658 structures collected from the Materials Project database [28, 29]
We have proposed a novel representation of crystalline materials named as ’orbital-field matrix (OFM)’, which is based on the distribution of valence shell electrons

Summary

Introduction

The increasing volume of available experimental and quantum-computational material data, along with the development of machine learning techniques, has provided a new opportunity to develop methods for accelerating discoveries of new materials and physical and chemical phenomena. By using machine learning algorithms, hidden information on materials, including patterns, features, chemical rules, and physical laws, can be automatically discovered from both firstprinciples-calculated data and experimental data [1,2,3,4,5,6,7,8]. From the viewpoint of data science, the material data using this primitive representation can be categorized as unstructured data, and the mathematical operations performed on such material data involve the algebra of sets only. Advanced quantitative machine learning algorithms cannot be applied directly and effectively to conventional material data, owing to the limitation of the algebraic operations of the primitive data representation. In order to apply well established machine learning methods, including predictive learning and descriptive learning, it is necessary to convert the

Methods

Results

Conclusion